Dependency Parsing Features for Semantic Parsing

Current semantic parsing systems either largely ignore the syntactic structure of the natural language input or attempt to learn highly underconstrained, noisy syntactic representations. We present two new classes of features for statistical semantic parsing that utilize information about syntactic structure, extracted from a dependency parse of an utterance, to score semantic parses for that utterance. These features, when added to an existing state-of-theart semantic parsing framework, improve macro-averaged F1 from 31% to 38.3% over a strong baseline on a broad-coverage benchmark dataset. In experiments on small, focused datasets, we identify specific relationships between syntactic structure and semantic composition that these features enable the framework to learn, relationships it fails to learn without them.